1. Hierarchical Series
In many applications, a set of time series is hierarchically organized. Examples include the presence of geographic levels, products, or categories that define different types of aggregations. In such scenarios, forecasters are often required to provide predictions for all disaggregate and aggregate series. A natural desire is for those predictions to be “coherent”, that is, for the bottom series to add up precisely to the forecasts of the aggregated series.

2. Hierarchical Forecast
To achieve “coherency”, most statistical solutions to the hierarchical forecasting challenge implement a two-stage reconciliation process.- First, we obtain a set of the base forecast
- Later, we reconcile them into coherent forecasts .
BottomUp
,
TopDown
,
MiddleOut
,
MinTrace
,
ERM
.
Among its probabilistic coherent methods we have
Normality
,
Bootstrap
,
PERMBU
.
3. Minimal Example
Wrangling Data
ds | top_level | middle_level | bottom_level | y | |
---|---|---|---|---|---|
0 | 2000-01-01 | Australia | State1 | r1 | 10 |
1 | 2000-02-01 | Australia | State1 | r1 | 20 |
8 | 2000-01-01 | Australia | State1 | r2 | 10 |
9 | 2000-02-01 | Australia | State1 | r2 | 20 |
16 | 2000-01-01 | Australia | State2 | r3 | 100 |
17 | 2000-02-01 | Australia | State2 | r3 | 200 |
24 | 2000-01-01 | Australia | State2 | r4 | 100 |
25 | 2000-02-01 | Australia | State2 | r4 | 200 |
Y_hier_df
dataframe.
The aggregation constraints matrix is captured
within the S_df
dataframe.
Finally the tags
contains a list within Y_hier_df
composing each
hierarchical level, for example the tags['top_level']
contains
Australia
’s aggregated series index.
unique_id | ds | y | |
---|---|---|---|
0 | Australia | 2000-01-01 | 220 |
1 | Australia | 2000-02-01 | 440 |
8 | Australia/State1 | 2000-01-01 | 20 |
9 | Australia/State1 | 2000-02-01 | 40 |
16 | Australia/State2 | 2000-01-01 | 200 |
17 | Australia/State2 | 2000-02-01 | 400 |
24 | Australia/State1/r1 | 2000-01-01 | 10 |
25 | Australia/State1/r1 | 2000-02-01 | 20 |
32 | Australia/State1/r2 | 2000-01-01 | 10 |
33 | Australia/State1/r2 | 2000-02-01 | 20 |
40 | Australia/State2/r3 | 2000-01-01 | 100 |
41 | Australia/State2/r3 | 2000-02-01 | 200 |
48 | Australia/State2/r4 | 2000-01-01 | 100 |
49 | Australia/State2/r4 | 2000-02-01 | 200 |
unique_id | Australia/State1/r1 | Australia/State1/r2 | Australia/State2/r3 | Australia/State2/r4 | |
---|---|---|---|---|---|
0 | Australia | 1.0 | 1.0 | 1.0 | 1.0 |
1 | Australia/State1 | 1.0 | 1.0 | 0.0 | 0.0 |
2 | Australia/State2 | 0.0 | 0.0 | 1.0 | 1.0 |
3 | Australia/State1/r1 | 1.0 | 0.0 | 0.0 | 0.0 |
4 | Australia/State1/r2 | 0.0 | 1.0 | 0.0 | 0.0 |
5 | Australia/State2/r3 | 0.0 | 0.0 | 1.0 | 0.0 |
6 | Australia/State2/r4 | 0.0 | 0.0 | 0.0 | 1.0 |
Base Predictions
Next, we compute the base forecast for each time series using thenaive
model. Observe that Y_hat_df
contains the forecasts but they
are not coherent.
Reconciliation
unique_id | ds | Naive | Naive/BottomUp | |
---|---|---|---|---|
0 | Australia | 2000-05-01 | 880.0 | 880.0 |
1 | Australia | 2000-06-01 | 880.0 | 880.0 |
4 | Australia/State1 | 2000-05-01 | 80.0 | 80.0 |
5 | Australia/State1 | 2000-06-01 | 80.0 | 80.0 |
8 | Australia/State2 | 2000-05-01 | 800.0 | 800.0 |
9 | Australia/State2 | 2000-06-01 | 800.0 | 800.0 |
12 | Australia/State1/r1 | 2000-05-01 | 40.0 | 40.0 |
13 | Australia/State1/r1 | 2000-06-01 | 40.0 | 40.0 |
16 | Australia/State1/r2 | 2000-05-01 | 40.0 | 40.0 |
17 | Australia/State1/r2 | 2000-06-01 | 40.0 | 40.0 |
20 | Australia/State2/r3 | 2000-05-01 | 400.0 | 400.0 |
21 | Australia/State2/r3 | 2000-06-01 | 400.0 | 400.0 |
24 | Australia/State2/r4 | 2000-05-01 | 400.0 | 400.0 |
25 | Australia/State2/r4 | 2000-06-01 | 400.0 | 400.0 |
References
- Hyndman, R.J., & Athanasopoulos, G. (2021). “Forecasting:
principles and practice, 3rd edition: Chapter 11: Forecasting
hierarchical and grouped series.”. OTexts: Melbourne, Australia.
OTexts.com/fpp3 Accessed on July
2022.
- Orcutt, G.H., Watts, H.W., & Edwards, J.B.(1968). Data aggregation
and information loss. The American Economic Review, 58 ,
773(787).
- Disaggregation methods to expedite product line forecasting.
Journal of Forecasting, 9 , 233–254.
doi:10.1002/for.3980090304.
- Wickramasuriya, S. L., Athanasopoulos, G., & Hyndman, R. J. (2019).
“Optimal forecast reconciliation for hierarchical and grouped time
series through trace minimization”. Journal of the American
Statistical Association, 114 , 804–819.
doi:10.1080/01621459.2018.1448825.
- Ben Taieb, S., & Koo, B. (2019). Regularized regression for
hierarchical forecasting without unbiasedness conditions. In
Proceedings of the 25th ACM SIGKDD International Conference on
Knowledge Discovery & Data Mining KDD ’19 (p. 1337(1347). New York,
NY, USA: Association for Computing
Machinery.